Data Stratification

Ensures the corpus reflects the relative volume and authority of different sovereign legal systems. A model pre-trained on a corpus dominated by U.S. case law will fail to reason about Civil Law systems.

• Common Law: U.S. federal/state, UK, Canada, Australia
• Civil Law: EU directives, German BGB, French Code Civil
• Hybrid Systems: Scotland, South Africa, Louisiana
• Supranational: CJEU, ECHR, WTO dispute rulings

Stratification prevents a model from applying common law stare decisis logic to a civil code jurisdiction where precedent is non-binding.

Balances the corpus across time periods to prevent temporal concept drift. A model trained only on recent data will lack historical legal context; one trained on outdated statutes will hallucinate repealed provisions.

• Historical: Foundational cases and superseded statutes for reasoning lineage
• Contemporary: Current enacted legislation and recent appellate decisions
• Emerging: Newly filed complaints, proposed regulations, and legal memos

This is critical for regulatory change detection and understanding the evolution of legal doctrine like the transformation of privacy torts into modern data protection law.

Controls the mix of legal practice areas to avoid domain overfitting. A corpus skewed toward securities filings will produce a model incapable of reasoning about criminal procedure or family law.

• Public Law: Constitutional, administrative, criminal, tax
• Private Law: Contracts, torts, property, corporate, IP
• Procedural Law: Rules of evidence, civil procedure, appellate rules
• Regulatory: Environmental, antitrust, financial services, healthcare

Each sub-domain has distinct linguistic registers and deontic logic structures that must be independently represented.

Stratifies by the structural and rhetorical category of legal text. A model must distinguish between the binding authority of a statute and the persuasive commentary of a law review article.

• Primary Authority: Statutes, constitutions, regulations, executive orders
• Secondary Authority: Restatements, uniform codes, model laws
• Adjudicatory: Judicial opinions, administrative decisions, arbitral awards
• Transactional: Contracts, deeds, wills, SEC filings, merger agreements
• Analytical: Legal briefs, memoranda, law review articles, treatises

Each type has a unique illocutionary force and argument structure that the model must learn to parse.

Ensures proportional representation of courts at different levels of the judicial hierarchy. A corpus dominated by trial court opinions will overfit to fact-finding language, while one with only supreme court decisions will miss procedural nuance.

• Trial Courts: District courts, circuit courts (original jurisdiction)
• Intermediate Appellate: Circuit Courts of Appeal, state appellate divisions
• Courts of Last Resort: Supreme courts, constitutional councils
• Specialized Courts: Tax Court, Federal Circuit, bankruptcy courts

This stratification teaches the model the weight of authority and the difference between holding and dicta at each level.

Measures and controls the interconnectedness of documents in the corpus via their citation graphs. A well-stratified corpus ensures that foundational cases with high in-degree centrality are adequately represented.

• Landmark Cases: High-citation nodes like Marbury v. Madison or Donoghue v. Stevenson
• Clusters: Tightly coupled citation networks in niche areas like patent obviousness
• Orphan Documents: Important but rarely cited administrative guidance

This property directly impacts a model's ability to perform citation network analysis and understand the precedential lineage of a legal proposition.